API Latest Releases: Java Core, C++ Core, Python, Memory, Pig, Hive,

Frequency Sketches Overview

Frequent Items

These sketches implement algorithms that are members of a class of “Heavy Hitters” algorithms that identify the “heaviest” or “most frequently occurring” items in a stream.

Suppose you have a web-site store that sells songs and you wish to identify the most frequent song-titles that are being downloaded from your store.

This is a perfect use-case for the frequencies/ItemsSketch, which is a Generic class that can be configured to count the number of occurrences of any arbitrary item. In this case our song-titles are strings. For example,

ItemsSketch<String> sketch = new ItemsSketch<String>();
while (remainingItems) { sketch.update("songTitle"); }

This configures the sketch to track and count frequent occurrences of Strings. And in this case you would update the sketch with the title of each song as it appears in the stream. Note that in this case we assume that each occurrence of a song title carries with it a “weight” of one. After the sketch has been populated with the stream you query the sketch to get a list of the “most frequently occurring” song titles with an approximate count of the actual number of occurences in the stream.

Now suppose your song titles are sold at different prices and you wish to identify the song titles that are generating the most revenue. In this case each item can carry a different “weight” which is the price. We can use the same sketch as before, but we update it using a “weight”.

ItemsSketch<String> sketch = new ItemsSketch<String>();
while (remainingItems) { sketch.update("songTitle", priceCents); }

The sketch only accepts integral values for the weight, so we just multiply the price by 100 to make the weight integer cents instead of fractional dollars.

The Frequent Items Sketch is an “aggregating” sketch in that duplicate items in the stream can have different weights and the sketch properly tracks the total weight for each distinct item.

Frequent Distinct Tuples Sketch

This is a very different algorithm that identifies the most frequent distinct occurrences associated with a specific key, and it is called the Frequent Distinct Tuples Sketch or FdtSketch. See the documentation for this sketch.